CN113310689B - Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network - Google Patents

Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network Download PDF

Info

Publication number
CN113310689B
CN113310689B CN202110588075.6A CN202110588075A CN113310689B CN 113310689 B CN113310689 B CN 113310689B CN 202110588075 A CN202110588075 A CN 202110588075A CN 113310689 B CN113310689 B CN 113310689B
Authority
CN
China
Prior art keywords
domain
data
target domain
source domain
source
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110588075.6A
Other languages
Chinese (zh)
Other versions
CN113310689A (en
Inventor
孙闯
李天福
赵志斌
王诗彬
田绍华
严如强
陈雪峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202110588075.6A priority Critical patent/CN113310689B/en
Publication of CN113310689A publication Critical patent/CN113310689A/en
Application granted granted Critical
Publication of CN113310689B publication Critical patent/CN113310689B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/028—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Acoustics & Sound (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention discloses a method for diagnosing faults of an aeroengine transmission system based on a domain self-adaptive graph convolution network, which comprises the following steps: acquiring first vibration signals of an aircraft engine transmission system with unknown faults at different rotating speeds, and taking the first vibration signals as label-free target domain data; constructing a domain self-adaptive graph convolution network, acquiring second vibration signals of an aircraft engine transmission system with known faults at different rotating speeds, taking the second vibration signals as source domain data with labels, and taking the source domain data and part of target domain data as input to train the domain self-adaptive graph convolution network; inputting the other part of target domain data into the trained domain self-adaptive graph convolution network to obtain a prediction label of the target domain data, and realizing the migration diagnosis of the faults of the transmission system of the aircraft engine at different rotating speeds.

Description

Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network
Technical Field
The disclosure belongs to the field of mechanical fault diagnosis, and particularly relates to a method for diagnosing faults of an aeroengine transmission system based on a domain self-adaptive graph convolution network.
Background
Intelligent diagnostics plays an important role in health management systems for aircraft engine transmission systems, which have been widely used in modern industry and whose primary purpose is to monitor equipment status and reduce down time. At present, an unsupervised domain self-adaptive method is successfully applied to mechanical fault diagnosis under variable working conditions. In the unsupervised domain self-adaptive method, three types of information such as class labels, domain labels and data structures are important in the process of realizing migration from a labeled source domain to an unlabeled target domain. However, most existing unsupervised domain adaptation methods only use the first two kinds of information, and neglect modeling of the data structure, which makes the information contained in the features extracted by the deep network incomplete. Therefore, there is a need for a model that can embed data structure information also in the extracted features.
The above information disclosed in this background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a domain-adaptive graph convolution network-based aeroengine transmission system fault diagnosis method, which converts the features extracted by a convolution neural network into association graph data through a design graph generation layer, and then uses the graph convolution network to model the generated association graph, so that the structural information among the data is also embedded into the extracted features, the information contained in the features is more complete, and the distinguishability of the extracted features and the robustness of a model are improved.
In order to achieve the above purpose, the present disclosure provides the following technical solutions:
a method for diagnosing faults of an aeroengine transmission system based on a domain self-adaptive graph convolution network comprises the following steps:
s100: acquiring first vibration signals of an aircraft engine transmission system with unknown faults at different rotating speeds, and taking the first vibration signals as label-free target domain data;
s200: constructing a domain self-adaptive graph convolution network, acquiring second vibration signals of an aircraft engine transmission system with known faults at different rotating speeds, taking the second vibration signals as source domain data with labels, and taking the source domain data and part of target domain data as input to train the domain self-adaptive graph convolution network;
s300: inputting the other part of target domain data into the trained domain self-adaptive graph convolution network to obtain a prediction label of the target domain data, and realizing the migration diagnosis of the faults of the transmission system of the aircraft engine at different rotating speeds.
Preferably, in step S200, the domain adaptive graph convolution network includes:
the characteristic extractor F comprises four layers of convolutional neural networks, one graph generation layer and two layers of graph convolution networks, and is used for extracting characteristic values of source domain data and part of target domain data;
the fault classifier C comprises a full connection layer and is used for identifying the fault type carried by the source domain data;
a domain discriminator D including three layers of map convolutional layers for discriminating a difference between source domain data and target domain data;
a data structure aligner S, said data structure aligner S comprising a maximum mean difference distance estimator for aligning data structures of source domain data and target domain data.
Preferably, the expression of the feature extractor F is:
Figure BDA0003087712100000031
and is
Figure BDA0003087712100000032
Wherein GConv (-) and Conv (-) denote graph convolution layer operation and convolution operation, respectively, GGL (-) denotes a graph generation layer,
Figure BDA0003087712100000033
x represents an adjacency matrix and a node characteristic matrix of the correlation diagram constructed by the diagram generation layer, respectively.
Preferably, the expression of the fault classifier C is:
C(X)＝aX+b
where a and b represent the weight and bias of the fully-connected layer, respectively.
Preferably, the expression of the domain discriminator D is:
D(Xs，Xt)＝Ｃonv(|Xs，Xt|)
wherein, XsAnd XtRespectively representing the learned source domain data features and target domain data features. | represents the splicing operation.
Preferably, the expression of the data structure aligner S is:
Figure BDA0003087712100000034
wherein the content of the first and second substances,
Figure BDA0003087712100000035
representing a set of mapping functions, k (·) representing
Figure BDA0003087712100000036
M and n represent the number of source domain data and target domain data, respectively,
Figure BDA0003087712100000037
and
Figure BDA0003087712100000038
respectively representing the ith sample from the source domain and the target domain,
Figure BDA0003087712100000039
and
Figure BDA00030877121000000310
respectively, the jth sample from the source domain and the target domain, s and t respectively, the source domain and the target domain, and i and j respectively, the position of the sample in the sample set.
Preferably, in step S200, the training of the domain adaptive graph convolution network includes the following steps:
s201: dividing the labeled source domain data into a first training sample and a first test sample, dividing part of unlabeled target domain data into a second training sample and a second test sample, and simultaneously inputting the first training sample and the second training sample into a feature extractor FThe convolutional neural network carries out feature extraction to obtain source domain features
Figure BDA0003087712100000041
And target domain characteristics
Figure BDA0003087712100000042
S202: characterizing the source domain
Figure BDA0003087712100000043
Inputting the data into a fault classifier C to obtain a predicted value of the fault category of the source domain data
Figure BDA0003087712100000044
And taking the cross entropy loss as a fault classification loss, wherein the fault classification loss is expressed as:
Figure BDA0003087712100000045
where E represents the mathematical expectation, CE represents the cross-entropy loss,
Figure BDA0003087712100000046
representing source domain samples
Figure BDA0003087712100000047
The label of (1);
s203: characterizing a source domain
Figure BDA0003087712100000048
And target domain characteristics
Figure BDA0003087712100000049
Inputting the prediction value of each type of data domain label into a domain discriminator D, recording the prediction value as 0 or 1, calculating the two-classification cross entropy loss with the real domain label after obtaining the domain label, and obtaining the domain classification loss, wherein the domain classification loss is expressed as:
Figure BDA0003087712100000051
wherein f () takes 0 or 1 to indicate whether this type of data belongs to the source domain or the target domain;
s204: characterizing a source domain
Figure BDA0003087712100000052
And target domain characteristics
Figure BDA0003087712100000053
Inputting into a data structure aligner S, measuring the structural difference between the source domain features and the target domain features through an MMD distance estimator to obtain a structural difference loss, wherein the structural difference loss is expressed as:
Figure BDA0003087712100000054
where φ (-) represents a non-linear mapping function, Ω represents a measure of this distance by embedding the extracted features into the regenerative kernel Hilbert space, | | - | computationally |2Representing a calculation of Euclidean distance;
s205: constructing an objective function using the fault classification loss, the domain classification loss, and the structural difference loss, the objective function being represented as:
LTotal＝LC+γLD+κLMMD
wherein L isTotalRepresenting the overall objective function, γ and κ representing the equilibrium coefficients;
s206: updating parameters of the domain adaptive graph convolution network according to the objective function, and repeatedly executing the steps S201 to S205 until reaching a specified training time, wherein the parameters of the domain adaptive graph convolution network are updated according to the following formula:
Figure BDA0003087712100000055
wherein the content of the first and second substances,
Figure BDA0003087712100000061
representing a differential operator, eta a learning rate, thetaF、θCAnd thetaDLearnable parameters, θ ', representing the feature extractor, the fault classifier, and the domain discriminator, respectively'CAnd θ'DRespectively representing parameters updated by learning, LTotal、LCAnd LDRespectively representing an overall objective function, a fault classification loss and a domain classification loss;
s207: and testing the trained domain adaptive graph convolution network by using the first test sample and the second test sample.
Preferably, the source domain characteristics are obtained
Figure BDA0003087712100000062
And target domain characteristics
Figure BDA0003087712100000063
Then, the extracted source domain features and target domain features need to be converted into associated graph data through a graph generation layer, and the specific conversion process comprises the following steps:
a. inputting the node characteristic matrix X extracted by the convolutional neural network into a full connection layer of a fault classifier C for nonlinear mapping;
b. calculating the product of the output of the full connection layer of the fault classifier C and the output transposition to obtain a product result matrix, and normalizing the matrix according to rows to obtain a normalized matrix A;
c. using Top-k sorting mechanism to select the Top k values with maximum value of each row in matrix A, thereby obtaining adjacent matrix
Figure BDA0003087712100000065
After the node characteristic matrix X and the adjacency matrix of the graph are obtained
Figure BDA0003087712100000066
The dependency graph data is obtained.
Preferably, the expression of the map-generating layer is:
Figure BDA0003087712100000064
wherein X represents a node feature matrix extracted by a convolutional neural network, MLP (-) represents a fully connected layer,
Figure BDA0003087712100000071
the output of the MLP is represented as,
Figure BDA0003087712100000072
to represent
Figure BDA0003087712100000073
Transpose of (a), normalization (b), A denotes the constructed adjacency matrix, Top-k (c) denotes the index returning the first k maxima of A line by line,
Figure BDA0003087712100000074
is the finally obtained adjacency matrix.
Preferably, after obtaining the association map data, modeling the association map data by using a graph convolution network so that the structural information between the source domain and the target domain data is embedded into the extracted source domain feature and the extracted target domain feature.
Compared with the prior art, the beneficial effect that this disclosure brought does:
the method and the device model class labels, domain labels and data structure information contained in unsupervised domain self-adaptation, so that the information contained in the extracted vibration signal features is more complete, the domain difference between source domain data and target domain data is reduced, and the distinguishability of the features and the diagnosis precision on the target domain data are improved.
Drawings
FIG. 1 is a flow chart of a method for diagnosing faults of an aeronautical engine transmission system based on a domain adaptive graph convolution network according to an embodiment of the present disclosure;
FIG. 2 is a schematic diagram of a domain adaptive graph convolution network provided by another embodiment of the present disclosure;
FIG. 3 is a schematic diagram of a process for generating a correlation diagram in a diagram generation layer according to another embodiment of the disclosure;
4(a) to 4(c) are schematic diagrams of correlation diagram data modeling provided by another embodiment of the present disclosure;
5(a) -5 (d) are vibration signals at different faults provided by another embodiment of the present disclosure;
FIG. 6 is a schematic diagram of the results of a fault migration diagnosis of an aircraft engine driveline under different operating conditions according to another embodiment of the present disclosure.
Detailed Description
Specific embodiments of the present disclosure will be described in detail below with reference to fig. 1 to 6 of the accompanying drawings. While specific embodiments of the disclosure are shown in the drawings, it should be understood that the disclosure can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
It should be noted that certain terms are used throughout the description and claims to refer to particular components. As one skilled in the art will appreciate, various names may be used to refer to a component. This specification and claims do not intend to distinguish between components that differ in name but not function. In the following description and in the claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. The description which follows is a preferred embodiment of the invention, but is made for the purpose of illustrating the general principles of the invention and not for the purpose of limiting the scope of the invention. The scope of the present disclosure is to be determined by the terms of the appended claims.
To facilitate an understanding of the embodiments of the present disclosure, the following detailed description is to be considered in conjunction with the accompanying drawings, and the drawings are not to be construed as limiting the embodiments of the present disclosure.
In one embodiment, as shown in fig. 1, a method for diagnosing a failure of an aeronautical engine transmission system based on a domain adaptive graph convolution network includes the following steps:
s100: collecting vibration signals of an aircraft engine transmission system at different rotating speeds;
in the step, bevel gear breakage, bevel gear abrasion, bevel gear return to factory overhaul and normal state of the transmission system of the aircraft engine at different rotating speeds are classified. Firstly, sample division is carried out on collected signals in four states, and corresponding labels are marked on the divided samples, namely 0 represents broken teeth, 1 represents gear wear, 2 represents factory return overhaul, and 3 represents normal. After the corresponding training samples are obtained, the domain adaptive graph convolution network disclosed by the disclosure is trained. And finally, inputting the test sample into a trained network, outputting a prediction result of the fault classifier C, if the output value is 0, indicating that the fault is broken teeth, if the output value is 1, indicating that the gear is worn, if the output value is 2, indicating that the gear is returned to a factory for overhaul, and if the output value is 3, indicating that the gear is in a normal state.
S200: constructing a domain self-adaptive graph convolution network, and training the domain self-adaptive graph convolution network by taking labeled source domain data and part of unlabeled target domain data as input data;
in this step, 20% of the target domain data is generally selected to train the domain adaptive graph convolution network, so that the feature extractor learns how to effectively extract features in the target domain data.
S300: and inputting the residual label-free target domain data into the trained domain self-adaptive graph convolution network, outputting the prediction result of the fault classifier, and realizing fault migration diagnosis of the aero-engine transmission system at different rotating speeds.
The above embodiments constitute a complete technical solution of the present disclosure, and by modeling the class label, the domain label and the data structure information included in the unsupervised domain adaptation, the information included in the extracted vibration signal feature is more complete, thereby reducing the domain difference between the source domain and the target domain data, and improving the distinguishability of the feature and the diagnosis precision on the target domain data.
In another embodiment, in step S200, as shown in fig. 2, the domain adaptive graph convolution network includes:
the characteristic extractor F comprises four layers of convolutional neural networks, one graph generation layer and two graph convolution layers, and is used for extracting a vibration signal characteristic value;
the fault classifier C comprises a full connection layer and is used for classifying and identifying faults;
a domain discriminator D including three layers of map convolutional layers for discriminating a difference between source domain data and target domain data;
a data structure aligner S, said data structure aligner S comprising a maximum mean difference distance estimator for aligning data structures of source domain data and target domain data.
The expression of the feature extractor F is:
Figure BDA0003087712100000101
and is
Figure BDA0003087712100000102
Wherein GConv (-) and Conv (-) denote graph convolution layer operation and convolution operation, respectively, GGL (-) denotes a graph generation layer,
Figure BDA0003087712100000103
x represents an adjacency matrix and a node feature matrix of the correlation diagram constructed by the diagram generation layer, respectively.
The expression of the fault classifier C is as follows:
C(X)＝aX+b
where a and b represent the weight and bias of the fully-connected layer, respectively.
The expression of the domain discriminator D is:
D(Xs，Xt)＝Ｃonv(|Xs，Xt|)
wherein, XsAnd XtRespectively representing the learned source domain data features and target domain data features. | represents the splicing operation.
The expression of the data structure aligner S is:
Figure BDA0003087712100000111
wherein the content of the first and second substances,
Figure BDA0003087712100000112
representing a set of mapping functions, k (-) representing
Figure BDA0003087712100000113
M and n represent the number of source domain data and target domain data, respectively,
Figure BDA0003087712100000114
and
Figure BDA0003087712100000115
respectively representing the ith sample from the source domain and the target domain,
Figure BDA0003087712100000116
and
Figure BDA0003087712100000117
respectively, j-th samples from the source domain and the target domain, s and t respectively, samples from the source domain and the target domain, and i and j respectively, the positions of the samples in the sample set.
In another embodiment, in step S200, training the domain adaptive graph convolution network includes the following steps:
s201: dividing the labeled source domain data into a first training sample and a first test sample, dividing part of unlabeled target domain data into a second training sample and a second test sample, and dividing the first training sample and the second training sampleSimultaneously inputting the convolution neural network in the feature extractor F for feature extraction to obtain source domain features
Figure BDA0003087712100000118
And target domain characteristics
Figure BDA0003087712100000119
S202: characterizing the source domain
Figure BDA00030877121000001110
Inputting the data into a fault classifier C to obtain a predicted value of the fault category of the source domain data
Figure BDA00030877121000001111
And taking the cross entropy loss as a fault classification loss, wherein the fault classification loss is expressed as:
Figure BDA0003087712100000121
where E represents the mathematical expectation, CE represents the cross-entropy loss,
Figure BDA0003087712100000122
representing source domain samples
Figure BDA0003087712100000123
The label of (1);
s203: characterizing a source domain
Figure BDA0003087712100000124
And target domain characteristics
Figure BDA0003087712100000125
Inputting the prediction value of each type of data domain label into a domain discriminator D, recording the prediction value as 0 or 1, calculating the two-classification cross entropy loss with the real domain label after obtaining the domain label, and obtaining the domain classification loss, wherein the domain classification loss is expressed as:
Figure BDA0003087712100000126
wherein Γ (·) takes a value of 0 or 1 to indicate whether the class of data belongs to a source domain or a target domain;
s204: characterizing a source domain
Figure BDA0003087712100000127
And target domain characteristics
Figure BDA0003087712100000128
Inputting into a data structure aligner S, measuring the structural difference between the source domain features and the target domain features through an MMD distance estimator to obtain a structural difference loss, wherein the structural difference loss is expressed as:
Figure BDA0003087712100000129
where φ (-) represents a non-linear mapping function, Ω represents a measure of this distance by embedding the extracted features into the regenerative kernel Hilbert space, | | - | computationally |2Representing a calculation of Euclidean distance;
s205: constructing an objective function using the fault classification loss, the domain classification loss, and the structural difference loss, the objective function being expressed as:
LTotal＝LC+γLD+κLMMD
wherein L isTotalRepresenting the overall objective function, γ and κ representing the equilibrium coefficients;
s206: updating parameters of the domain adaptive graph convolution network according to the objective function, and repeatedly executing the steps S201 to S205 until reaching a specified training time, wherein the parameters of the domain adaptive graph convolution network are updated according to the following formula:
Figure BDA0003087712100000131
wherein the content of the first and second substances,
Figure BDA0003087712100000132
representing a differential operator, eta a learning rate, thetaF、θＣAnd thetaDLearnable parameters, θ ', representing the feature extractor, the fault classifier, and the domain discriminator, respectively'CAnd θ'DRespectively representing parameters updated by learning, LTotal、LCAnd LDRespectively representing the overall objective function, the fault classification loss and the domain classification loss.
In this step, a parameter θ is setFIs 1, the learning rate is 0.001, and the parameter theta is corrected according to the overall objective functionFObtaining a deviation derivative
Figure BDA0003087712100000133
With a value of 3, θ can be obtained according to the above formulaFThe updated value was 0.997.
S207: and testing the trained domain adaptive graph convolution network by using the first test sample and the second test sample.
In the step, the performance of the domain self-adaptive graph convolution network is evaluated through a total precision index, and the evaluation method comprises the following steps: the total number of correctly classified samples is divided by the total number of samples.
If the overall accuracy index is lower than the set value in the test process, the number of training samples needs to be increased to train the model again until the index reaches the set value.
In another embodiment, after obtaining the source domain features and the target domain features, the extracted source domain features and target domain features need to be converted into association map data through a map generation layer, as shown in fig. 3, a specific conversion process includes the following steps:
a. inputting the node characteristic matrix X extracted by the convolutional neural network into a full connection layer for nonlinear mapping;
b. calculating the product of the output of the full-connection layer and the output transpose to obtain a product result matrix, and normalizing the matrix according to rows to obtain a normalized matrix A;
c. using Top-k sorting mechanism to select the Top k values with maximum value of each row in matrix A, thereby obtaining adjacent matrix
Figure BDA0003087712100000141
After the node characteristic matrix X and the adjacency matrix of the graph are obtained
Figure BDA0003087712100000142
The dependency graph data is obtained.
In this embodiment, the expression of the map generation layer is:
Figure BDA0003087712100000143
wherein X represents a node feature matrix extracted by a convolutional neural network, MLP (-) represents a fully connected layer,
Figure BDA0003087712100000144
and represents the output of the MLP (or MLP),
Figure BDA0003087712100000145
to represent
Figure BDA0003087712100000146
Normalized, A represents the constructed adjacency matrix, Top-k (-) represents the index that returns the first k maxima of A row by row,
Figure BDA0003087712100000147
is the finally obtained adjacency matrix.
In another embodiment, after obtaining the association graph data, modeling the association graph data through a graph convolution network, so that the structural information between the source domain data and the target domain data is embedded into the extracted source domain features and target domain features.
In this embodiment, as shown in fig. 4(a) to 4(c), modeling the association graph data by the graph convolution network includes the following steps:
a. as shown in fig. 4(a), determining the range of the aggregation neighborhood node at each convolution, where K ═ 1 indicates that the range of the aggregation neighborhood node is a node at a distance of 1 from the central node;
b. as shown in fig. 4(b), each node iterates to aggregate node information with a distance of 1 until a specified number of training times is reached;
c. as shown in fig. 4(c), the node feature output is learned and can be used for node classification and fault diagnosis.
In one embodiment, the acquired vibration signals at 4 rotation speeds are respectively numbered as 0, 1, 2 and 3, wherein, as shown in fig. 5(a), 0 represents a normal state; as shown in fig. 5(b), 1 represents surface wear; as shown in fig. 5(c), 2 denotes a broken tooth; the tip end collapse is shown at 3 in FIG. 5 (d). One group of data collected under any two different working conditions is used as source domain data and is labeled, and the other group of data is used as target domain data without labels, so that 12 migration diagnosis tasks can be performed in total. In the experiment, 5 existing methods, namely, a Convolutional Neural Network (CNN), a maximum mean difference method (MMD), a relationship alignment method (CORAL), a domain adaptive network (DANN), and a conditional domain adaptive network (CDANN), were compared in addition to the Domain Adaptive Graph Convolution Network (DAGCN), and the experimental results are shown in table 1.
TABLE 1
Figure BDA0003087712100000151
Figure BDA0003087712100000161
FIG. 6 is a schematic illustration of the aeroengine driveline fault migration diagnostic results of the present disclosure at different rotational speeds.
From the results of table 1 and fig. 6, it can be seen that the average value of the diagnosis results obtained by the domain adaptive graph convolution network in each migration task is 78.81%, which is better than all the comparison methods, and the domain adaptive graph convolution network can obtain the best diagnosis effect in each migration task. The model can learn the characteristics with the domain invariant characteristics, so that the effective migration diagnosis of the aircraft engine transmission system under different working conditions is realized.
The foregoing describes the general principles of the present disclosure in conjunction with specific embodiments, however, it is noted that the advantages, effects, etc. mentioned in the present disclosure are merely examples and are not limiting, and they should not be considered essential to the various embodiments of the present disclosure. Furthermore, the foregoing disclosure of specific details is for the purpose of illustration and description and is not intended to be limiting, since the foregoing disclosure is not intended to be exhaustive or to limit the disclosure to the precise details disclosed.

Claims (9)

1. A method for diagnosing faults of an aeroengine transmission system based on a domain self-adaptive graph convolution network comprises the following steps:
s100: acquiring first vibration signals of an aircraft engine transmission system with unknown faults at different rotating speeds, and taking the first vibration signals as label-free target domain data;
s200: constructing a domain self-adaptive graph convolution network, acquiring a second vibration signal of an aircraft engine transmission system with known faults at different rotating speeds, taking the second vibration signal as source domain data with labels, and taking the source domain data and part of target domain data as input to train the domain self-adaptive graph convolution network, wherein the domain self-adaptive graph convolution network comprises:
the characteristic extractor F comprises four layers of convolutional neural networks, a graph generation layer and two layers of graph convolutional networks, is used for extracting characteristic values of source domain data and part of target domain data, and is also used for converting the extracted characteristic values of the source domain data and the target domain data into association graph data and modeling the association graph data;
the fault classifier C comprises a full connection layer and is used for identifying the fault type carried by the source domain data;
a domain discriminator D including three layers of map convolutional layers for discriminating a difference between source domain data and target domain data;
a data structure aligner S, said data structure aligner S comprising a maximum mean difference distance estimator for aligning data structures of source domain data and target domain data;
s300: and inputting the other part of target domain data into the trained domain self-adaptive graph convolution network to obtain a prediction label of the target domain data, thereby realizing the migration diagnosis of the faults of the transmission system of the aero-engine at different rotating speeds.
2. The method of claim 1, wherein the feature extractor F is expressed as:
Figure FDA0003537183470000011
and is
Figure FDA0003537183470000021
Wherein GConv (-) and Conv (-) denote graph convolution layer operation and convolution operation, respectively, GGL (-) denotes a graph generation layer,
Figure FDA0003537183470000022
x represents an adjacency matrix and a node feature matrix of the correlation diagram constructed by the diagram generation layer, respectively.
3. The method of claim 1, wherein the fault classifier C has the expression:
C(X)＝aX+b
where a and b represent the weight and bias of the fully-connected layer, respectively.
4. The method of claim 1, wherein the domain discriminator D has the expression:
D(Xs,Xt)＝Conv(|Xs,Xt|)
wherein, XsAnd XtRespectively representing the learned source domain data characteristics and target domain data characteristics, |, representing the splicing operation.
5. The method of claim 1, wherein the data structure aligner, S, is expressed as:
Figure FDA0003537183470000023
wherein the content of the first and second substances,
Figure FDA0003537183470000024
representing a set of mapping functions, k (·) representing
Figure FDA0003537183470000027
M and n represent the number of source domain data and target domain data, respectively,
Figure FDA0003537183470000025
and
Figure FDA0003537183470000026
respectively representing the ith sample from the source domain and the target domain,
Figure FDA0003537183470000031
and
Figure FDA0003537183470000032
respectively, j-th samples from the source domain and the target domain, s and t respectively, samples from the source domain and the target domain, and i and j respectively, the positions of the samples in the sample set.
6. The method of claim 1, wherein the training of the domain adaptive graph convolution network in step S200 comprises the steps of:
s201: dividing source domain data with labels into a first training sample and a first test sample, dividing part of target domain data without labels into a second training sample and a second test sample, simultaneously inputting the first training sample and the second training sample into a convolutional neural network in a feature extractor F for feature extraction to obtain source domain features
Figure FDA0003537183470000033
And target domain characteristics
Figure FDA0003537183470000034
S202: characterizing the source domain
Figure FDA0003537183470000035
Inputting the data into a fault classifier C to obtain a predicted value of a source domain data fault category
Figure FDA0003537183470000036
And taking the cross entropy loss as a fault classification loss, wherein the fault classification loss is expressed as:
Figure FDA0003537183470000037
where E represents the mathematical expectation, CE represents the cross-entropy loss,
Figure FDA0003537183470000038
representing source domain samples
Figure FDA0003537183470000039
The label of (1);
s203: characterizing a source domain
Figure FDA00035371834700000310
And target domain characteristics
Figure FDA00035371834700000311
Inputting the prediction value of each type of data domain label into a domain discriminator D, recording the prediction value as 0 or 1, calculating the two-classification cross entropy loss with the real domain label after obtaining the domain label, and obtaining the domain classification loss, wherein the domain classification loss is expressed as:
Figure FDA0003537183470000041
wherein Γ (·) takes either 0 or 1 to indicate whether the class of data belongs to the source domain or the target domain;
s204: characterizing a source domain
Figure FDA0003537183470000042
And target domain characteristics
Figure FDA0003537183470000043
Inputting into a data structure aligner S, measuring the structural difference between the source domain features and the target domain features through an MMD distance estimator to obtain a structural difference loss, wherein the structural difference loss is expressed as:
Figure FDA0003537183470000044
where φ (-) represents a non-linear mapping function, Ω represents a measure of this distance by embedding the extracted features into the regenerative kernel Hilbert space, | | - | computationally |2Representing a calculation of Euclidean distance;
s205: constructing an objective function using the fault classification loss, the domain classification loss, and the structural difference loss, the objective function being represented as:
LTotal＝LC+γLD+κLMMD
wherein L isTotalRepresents the overall objective function, γ and κ represent the balance coefficients;
s206: updating parameters of the domain adaptive graph convolution network according to the objective function, and repeatedly executing the steps S201 to S205 until reaching a specified training time, wherein the parameters of the domain adaptive graph convolution network are updated according to the following formula:
Figure FDA0003537183470000051
wherein the content of the first and second substances,
Figure FDA0003537183470000052
representing a differential operator, eta a learning rate, thetaF、θCAnd thetaDLearnable parameters, θ ', of the feature extractor, the fault classifier and the domain discriminator, respectively'CAnd θ'DRespectively representing parameters updated by learning, LTotal、LCAnd LDRespectively representing an overall objective function, a fault classification loss and a domain classification loss;
s207: and testing the trained domain adaptive graph convolution network by using the first test sample and the second test sample.
7. The method of claim 6, wherein the source domain characteristics are obtained
Figure FDA0003537183470000053
And target domain characteristics
Figure FDA0003537183470000054
Then, the extracted source domain features and target domain features need to be converted into associated graph data through a graph generation layer, and the specific conversion process comprises the following steps:
a. inputting the node characteristic matrix X extracted by the convolutional neural network into a full connection layer of a fault classifier C for nonlinear mapping;
b. calculating the product of the output of the full-connection layer of the fault classifier C and the output transpose to obtain a product result matrix, and normalizing the matrix according to rows to obtain a normalized matrix A;
c. using Top-k sorting mechanism to select the Top k values with maximum value of each row in matrix A, thereby obtaining adjacent matrix
Figure FDA0003537183470000056
After the node characteristic matrix X and the adjacency matrix of the graph are obtained
Figure FDA0003537183470000055
The dependency graph data is obtained.
8. The method of claim 7, wherein the graph generation layer is expressed as:
Figure FDA0003537183470000061
wherein X represents a node feature matrix extracted by a convolutional neural network, MLP (-) represents a fully connected layer,
Figure FDA0003537183470000062
the output of the MLP is represented as,
Figure FDA0003537183470000063
to represent
Figure FDA0003537183470000064
Transpose of (a), normaize (-) denotes normalization, A denotes the constructed adjacency matrix, Top-k (-) denotes the index returning the first k maxima of A line by line,
Figure FDA0003537183470000065
is the finally obtained adjacency matrix.
9. The method of claim 7, wherein after obtaining the association graph data, the association graph data is modeled by using a graph convolution network such that structural information between the source domain and the target domain data is embedded in the extracted source domain features and target domain features.
CN202110588075.6A 2021-05-27 2021-05-27 Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network Active CN113310689B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110588075.6A CN113310689B (en) 2021-05-27 2021-05-27 Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110588075.6A CN113310689B (en) 2021-05-27 2021-05-27 Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network

Publications (2)

Publication Number Publication Date
CN113310689A CN113310689A (en) 2021-08-27
CN113310689B true CN113310689B (en) 2022-07-12

Family

ID=77375725

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110588075.6A Active CN113310689B (en) 2021-05-27 2021-05-27 Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network

Country Status (1)

Country Link
CN (1) CN113310689B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114048546B (en) * 2021-11-17 2024-04-23 大连理工大学 Method for predicting residual service life of aeroengine based on graph convolution network and unsupervised domain self-adaption
CN114969998A (en) * 2022-03-08 2022-08-30 西安交通大学 Engine fault diagnosis method based on fault feature migration and progressive fine adjustment

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110210381B (en) * 2019-05-30 2023-08-25 盐城工学院 Domain separation self-adaptive one-dimensional convolutional neural network intelligent fault diagnosis method
CN111353373B (en) * 2019-11-29 2023-08-04 盐城工学院 Related alignment domain adaptive fault diagnosis method
CN111898095B (en) * 2020-07-10 2024-04-19 佛山科学技术学院 Deep migration learning intelligent fault diagnosis method, device, storage medium and equipment
CN112183581A (en) * 2020-09-07 2021-01-05 华南理工大学 Semi-supervised mechanical fault diagnosis method based on self-adaptive migration neural network

Also Published As

Publication number Publication date
CN113310689A (en) 2021-08-27

Similar Documents

Publication Publication Date Title
CN114048568B (en) Rotary machine fault diagnosis method based on multisource migration fusion shrinkage framework
CN113935406B (en) Mechanical equipment unsupervised fault diagnosis method based on countermeasure flow model
CN111914883B (en) Spindle bearing state evaluation method and device based on deep fusion network
CN113310689B (en) Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network
CN113375941B (en) Open-set fault diagnosis method for bearings of high-speed motor train unit
CN115600150A (en) Multi-mode gearbox fault diagnosis method based on deep migration learning
CN116894187A (en) Gear box fault diagnosis method based on deep migration learning
CN114297918A (en) Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning
CN115791174B (en) Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium
WO2019178930A1 (en) Fault diagnosis method for mechanical device
CN115859077A (en) Multi-feature fusion motor small sample fault diagnosis method under variable working conditions
CN114548199A (en) Multi-sensor data fusion method based on deep migration network
CN113239610A (en) Domain self-adaptive rolling bearing fault diagnosis method based on Wasserstein distance
CN116011507A (en) Rare fault diagnosis method for fusion element learning and graph neural network
CN114564987A (en) Rotary machine fault diagnosis method and system based on graph data
CN115563565A (en) Motor bearing fault diagnosis method and device based on incremental learning
CN116625686A (en) On-line diagnosis method for bearing faults of aero-engine
CN115757103A (en) Neural network test case generation method based on tree structure
CN111783941B (en) Mechanical equipment diagnosis classification method based on probability confidence convolutional neural network
CN116894215B (en) Gear box fault diagnosis method based on semi-supervised dynamic graph attention
CN112163630A (en) Compound fault diagnosis method and device based on unbalanced learning
CN116644348A (en) Cross-mechanical part fault diagnosis method and device based on transfer type countermeasure migration
CN116644273A (en) Fault diagnosis method and system based on interpretability multiplication convolution network
CN113778811A (en) Fault monitoring method and system based on deep convolution migration learning software system
CN113723592A (en) Fault diagnosis method based on wind power gear box monitoring system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant